A coding agent that bet on bring-your-own-model
Cline is an AI agent that edits your code, runs your terminal, and uses tools, with a human in the loop on every step by default. The thing that sets it apart from a closed assistant like Copilot or Cursor is the wiring: Cline never sits between you and the model. You plug in your own key (Anthropic, OpenAI, Gemini, OpenRouter, Bedrock, or a local Ollama endpoint) and pay the provider directly, with no markup and no proxy. That single decision shapes everything else about how it behaves and what it costs.
It ships as a VS Code extension, a JetBrains plugin, and a standalone CLI, all built on the same agent core (@cline/sdk). The repo crossed 63,000 stars as of 2026-06 and pushes releases multiple times a week, which is the kind of cadence that matters for a tool whose job is to keep up with new models.
What it actually does
The core loop is read, edit, run, observe. Cline reads your project structure and the relationships between files, makes coordinated edits across them, and watches linter and compiler output as it works, so it can catch a missing import or a type mismatch before you do. Every edit lands as a reviewable diff, and every step is tracked with a checkpoint you can roll back to. That checkpoint model is the practical difference from a terminal tool that only has git history to fall back on.
Beyond editing, it runs bash commands and reacts to their output in real time, connects to MCP servers for databases and APIs, and reads project rules from .clinerules files so it follows your conventions instead of inventing its own.
Install
Cline is not one package. Pick the surface you want:
# CLI (interactive or headless for CI)
npm i -g cline
# Programmatic agents and integrations
npm install @cline/sdkFor the editors, install from the marketplaces rather than npm: search for Cline in the VS Code Marketplace (the extension id is saoudrizwan.claude-dev), or install the Cline plugin from the JetBrains Marketplace for IntelliJ, PyCharm, WebStorm, GoLand, and the rest of the family. All of them talk to the same agent core, so a workflow you build in one carries over.
Plan mode and Act mode
The feature worth understanding before you start is the two-mode split. In Plan mode Cline explores the codebase, asks clarifying questions, and proposes a strategy without touching anything. Once you agree, you switch to Act mode and it executes. Each file edit and command still asks for approval unless you turn on auto-approve. This is Cline’s answer to the usual failure of autonomous agents, which is confidently doing the wrong thing across twenty files. Planning first, then acting, gives you a checkpoint to reject before any damage is done.
Where it fits, and where it doesn’t
Reach for Cline when you want an IDE-native agent with full model choice and a visible diff for every change, when your team needs to standardize behavior through rules and skills, or when you want to script it headless in CI with JSON output. It is a good fit for developers who already pay for model API access and want the agent layer to be transparent about cost.
It is a worse fit if you want a fixed monthly price with no per-token accounting, since you are billed by your provider for whatever the agent consumes. It is also not the lightest option for quick single-file edits, where an inline assistant has less ceremony.
cline versus the other coding agents
| cline | Aider | Continue | OpenHands | |
|---|---|---|---|---|
| Stars | 63,353 | 46,274 | 33,704 | 77,254 |
| Surface | VS Code, JetBrains, CLI | terminal | VS Code, JetBrains | Docker app, CLI, cloud |
| Shape | IDE agent with approval gates | git-native pair programmer | embedded assistant plus autocomplete | sandboxed autonomous agent |
| Model access | bring your own key | bring your own key | bring your own key | bring your own key |
| Best at | reviewed edits across files | commit-per-change workflows | inline help in the editor | hands-off long tasks in a sandbox |
Counts are from GitHub as of June 2026, all under Apache-2.0 except OpenHands. Aider lives in the terminal and commits every change to git, which suits a different review style. Continue leans toward autocomplete and inline chat rather than a full agent loop. OpenHands runs the agent inside a Docker sandbox for fully hands-off work, at the cost of heavier setup. Cline sits in the middle: an agent that lives in your editor and shows its work.
The star curve
The star history rises steadily rather than in one viral spike, which tracks with how the project grew through the 2025 wave of coding agents and kept compounding as it added the CLI, SDK, and JetBrains support. There is no single launch moment to point at; the growth is the story.
What the issue tracker warns you about
Two recurring themes in the tracker are worth knowing before you commit. First, token cost can surprise you. Because Cline sends substantial file context and sometimes retries when a diff edit fails to apply cleanly, real spend can run higher than a back-of-envelope “one model call” estimate, and several issues describe duplicate requests or diff mismatches inflating bills. Budget for it, and watch your provider dashboard the first few sessions.
Second, auto-approve is genuinely risky. Reports exist of the agent removing or rewriting files during cleanup steps when running unattended, so treat full auto-approve as something to enable only on a branch you can throw away. Model compatibility is the third caveat: tool use through some local models and some third-party gateways has been flaky, so the smoothest path is a frontier model from a first-party provider.
Related
For a terminal-first alternative, see Aider. For an editor assistant with autocomplete, see Continue. For a sandboxed autonomous agent, see OpenHands. Another open-source agent worth a look is opencode. For what else is climbing, see LLM tooling, the daily digest, and the weekly report.
FAQ
Is Cline free? The software is open source under Apache-2.0, so there is no license fee. You pay your model provider directly for token usage, because Cline adds no markup and runs no proxy.
Cline vs Aider, which should I use? Aider runs in the terminal and commits every change to git, which fits a CLI-centric, review-by-diff workflow. Cline runs in your IDE with approval gates and rollback checkpoints. Choose Aider for git-native terminal work, Cline for editor-native agent work.
Can Cline run fully offline with a local model? Partly. You can point it at a local Ollama or LM Studio model, but tool use with some local models has been unreliable, and any MCP servers you connect for databases or APIs still need network access. Test your specific model before relying on it.
Should I turn on auto-approve in production? Not on code you cannot afford to lose. The issue tracker documents cases of unattended runs deleting or rewriting files during cleanup, so keep auto-approve to disposable branches and review diffs otherwise.
How is Cline related to Roo Code? Roo Code began as a fork of Cline and has since followed its own roadmap. They share lineage but are now separate projects with different teams and feature sets.